#011: AI Needs Better Metaphors - by Mete Polat
SubscribeSign in
#011: AI Needs Better Metaphors<br>Plus, my agent pays the bills, personal software goes native, and the value of being unemployed.
Mete Polat<br>May 05, 2026
Share
Imagine the scene. It’s Thanksgiving 2026 and you’re sitting next to your uncle at the dinner table. After a questionable rant about how prediction markets are the oracles of truth in the 21st century, he tells you he’s thinking about “getting into AI.” You’re relieved to finally switch the topic. He asks if you could help explain some basics. Easy-peasy, you think - you’re so AI-pilled that random numbers like 4.7 and 5.5 evoke emotional responses.<br>“So what’s the difference between a skill and a plugin?” he asks. “What about MCP? Is that the same as CLI? And is that the same as API?” You start sweating. “Wait, are sub-agents like agents but they’re worse than normal agents? Do CLI agents work for the CLI? What’s the difference between a routine, a scheduled task, and a cron job? Is the scheduled task not routine? What does cron mean - is that, like, the last name of the guy who invented it?” You are now questioning your life and starting to think the Polymarket sermon was actually the good part of dinner.<br>You think I’m joking, but please explain to me the difference between Notion’s custom agents, Claude Cowork’s scheduled tasks, and Claude Code’s routines:
Claude Cowork “scheduled tasks” on the left, Notion “custom agents” on the right.
Claude Code “routines”. Is it different from “scheduled tasks” in cowork? Good question. I have no idea.<br>At the risk of tooting my own horn, I’d say I’m as AI-pilled as they come in non-technical circles - I use AI for hours daily, I’ve built software with it, I built my own agents, skills, and plugins, I have a Hermes agent running on my Mac mini, and I have a postcard on my desk with a quote from a Claude Code session that made me laugh. And I still struggle to understand, let alone explain, the difference between all these concepts.<br>The problem isn’t just the confusing terminology. All these things can be understood and explained, but that’s not the point. The point is that we don’t yet have a solid set of well-fitting metaphors to ground these AI primitives. We don’t need better terminology or even better interfaces - first, we need better metaphors.<br>Cliff Kuang, in his book “User Friendly”1, lays out a convincing argument for why metaphors are foundational to broad technology adoption. I’m including two quotes here because they’ll do a much better job of explaining it than I will, so bear with me:<br>The right metaphor is like an instruction manual but better, because it teaches you how something should work without you ever having to be told. Consider the metaphor of the in-box versus the news feed. The email in-box borrows its logic from your mail, and you probably at least glance at every piece of mail that’s sent to you, simply because they were all meant for you. Your email in-box carries the same logic. The Instagram “feed” or the Twitter “stream” are entirely different metaphors. A stream rushes on even if you’re not there to see; it gurgles by in the dark, when you’re asleep. To say that information is a stream suggests that it’s there for the taking, if we wish to drink, not that we have to consume it all. A stream or a news feed, even if it’s crafted to your whims, doesn’t require your personal attention. It’s a commons to be shared.
and:<br>That’s how metaphors work: Once their underlying logic becomes manifest, we forget that they were ever there. No one remembers that before the steering wheel in a car, there were tillers, and that tillers made for a natural comparison when no one drove cars and far more people had piloted a boat. The metaphor disappeared once driving cars became common. In digesting new technologies, we climb a ladder of metaphors, and each rung helps us step up to the next.
Arguably, we’re already on the right path. We have collectively landed on one of the core metaphors that will underpin the next age of AI - AI as co-worker / employee. This is why the chat interface feels natural and is becoming pervasive in AI apps - we’ve been “chatting” with other humans through a computer for 4 decades. When you give something a name, a perfect command of language, and the same tools you have, it’s no different from your colleague Jerry who keeps promising to send you that brief that’s 2 weeks overdue (be honest, you can’t wait for Jerry to be replaced with AI).<br>The design challenge will be climbing this ladder further. The only primitives that seem to do that well currently are skills and memory - you can just understand what they are from the name and apply that mental model without problem to AI. Everything else starts getting hazy (and lazy) from here as engineers are trying to adapt terms from the previous waves of computing to describe these new concepts - cron jobs, MCPs, sub-agents, slash commands, plugins... you...